AI Job Screeners Favor AI-Written Resumes Over Human Ones, Researchers Find
New research suggests AI-based hiring systems can prefer resumes written by the same kind of model used for screening, raising fairness concerns for applicants and employers.
AI Job Screeners Favor AI-Written Resumes Over Human Ones, Researchers Find
A new concern in automated hiring
A growing body of hiring technology is facing fresh scrutiny after researchers found that AI resume screeners can systematically prefer resumes written by AI over those written by people. The core concern is not simply that automation is being used in recruiting, but that the same class of tools may now be shaping both sides of the process: applicants use large language models to polish resumes, while employers use similar models to evaluate them. According to the research, that overlap can create a structural advantage for candidates whose applications resemble the model’s own writing style.
“LLMs consistently prefer resumes generated by themselves.”
How the researchers tested it
The work was conducted by Jiannan Xu of the University of Maryland, Gujie Li of the National University of Singapore, and Jane Yi Jiang of Ohio State University. Using a dataset of 2,245 human-written resumes collected before generative AI became widespread, the researchers created matched AI-generated versions with models including GPT-4o, GPT-4-turbo, GPT-4o-mini, LLaMA 3.3-70B, Mistral-7B, Qwen 2.5-72B, and DeepSeek-V3. To isolate the effect of AI writing style, they kept factual sections such as work history, skills, and education the same, and focused on replacing only the more subjective executive summary portion.
That design mattered because it meant the comparison was not really between stronger and weaker candidates. Instead, it was a test of whether screening systems favored wording and structure that mirrored their own outputs. In the paper’s abstract, the authors said the bias against human-written resumes ranged from 67% to 82% across major models, and in simulated hiring pipelines across 24 occupations, candidates using the same model as the evaluator were 23% to 60% more likely to be shortlisted than equally qualified applicants with human-written resumes.
Where the effect appeared strongest
The findings suggest the impact was not uniform across all occupations. The simulations showed the biggest disadvantages for human-written resumes in business-related roles such as sales, accounting, and finance. The effect was described as less pronounced in areas such as agriculture, arts, and automotive, though still part of the broader pattern of model self-preference. Researchers warned that, over time, this could create a kind of lock-in effect, where the dominant stylistic patterns of major AI systems become embedded in applicant pools and influence who gets noticed.
That makes the issue especially important for the wider technology and employment landscape. If resume screening becomes a contest over matching the preferences of a particular model, hiring decisions may drift away from evaluating actual qualifications and toward rewarding access to the “right” AI tool. The authors framed that as a new type of fairness problem: not the traditional demographic bias often discussed in AI governance, but bias created through AI-to-AI interaction.
Possible fixes for employers
The research did not present the bias as unavoidable. The authors tested two relatively simple mitigation strategies: using system prompts that instruct the model to ignore where the text came from and focus on substance, and using a majority-voting ensemble so one model’s preferences are diluted by others. In the current arXiv version, the authors said these interventions reduced bias by more than 50%; the AIES publication abstract describes reductions of over 60%, reflecting somewhat different reported figures across versions of the work. In either case, the direction is clear: the bias can be meaningfully reduced without rebuilding hiring systems from scratch.
Why it matters beyond the study
For California’s Central Valley, the research does not center on a specific local employer or city, but the implications are easy to see. Employers across the region increasingly rely on digital tools to sort large pools of applicants, and job seekers in sectors ranging from office work to logistics, healthcare, and agriculture may feel pressure to use AI simply to avoid being filtered out. Even where the simulated effect appeared smaller in agriculture-related roles, the broader lesson is that automated screening may be rewarding stylistic compatibility with a model rather than human judgment about experience and fit.
More broadly, the findings matter for AI governance because they point to a subtle but consequential risk in modern software design. The issue is not only whether AI can replace repetitive recruiting tasks, but whether it quietly reshapes labor markets by favoring applicants who know how to write for machines. That makes resume screening a clear example of how technology systems can influence access to opportunity, even when the applicants themselves are similarly qualified.
Central Valley AI is produced by the CVAI Business Desk team and developed by Kaweah Tech, a regional firm that builds, deploys, and integrates AI solutions for businesses across California's Central Valley.